[Confirmed]Overclocking HBM performance test.

[dorsetknob]

Now, if this in some crazy way offered like a 20% boost in performance, everyone would be interested.

And everyone ( apart from the Green Team) would be thanking you

But as it stands the performance gains demonstrated by your example's DO NOT MATCH THE HYPE

 

[15th Warlock]

Just as I suspected, the gains aren't anywhere near the big jump in performance alleged to in the previous thread, and more in line with what you would expect from OCing memory on any card.

A bit off topic, but related nontheless, I was just done reading anandtech's excellent review of the Fury, it's an excellent read and I recommend it, after reading it I have a better grasp of how AMD over engineered the Fury, everything from the 500W cooling solution, to the amazing feat of developing the right tools to bring HBM to the masses, seems like AMD threw all of its weight on the Fury, and they can only be commended for doing so.

The Fury is not a bad card by any stretch of the imagination, the only problem is, the 980Ti really upset the balance of power before the Fury was even launched, as btruner suggested in his excellent editorial opinion article.

The potential to OC the card with such an excellent cooling solution is there, the only thing needed is for tools to be developed to unlock voltage control, then this little card could OC to new heights, dunno if it'll OC as well as Maxwell does, but it will surely go beyond the paltry 50MHz or 75MHz most users and reviewers are getting currently.

That's the untapped potential in Fury, HBM OCing is cool and all that, but as it stands, the card is not starving for memory bandwidth, but that core sure can and will be pushed to the limits by extreme OCers once they get their hands on the right tools, then things will get interesting

 

[the54thvoid]

Just as I suspected, the gains aren't anywhere near the big jump in performance alleged to in the previous thread, and more in line with what you would expect from OCing memory on any card.

A bit off topic, but related nontheless, I was just done reading anandtech's excellent review of the Fury, it's an excellent read and I recommend it, after reading it I have a better grasp of how AMD over engineered the Fury, everything from the 500W cooling solution, to the amazing feat of developing the right tools to bring HBM to the masses, seems like AMD threw all of its weight on the Fury, and they can only be commended for doing so.

The Fury is not a bad card by any stretch of the imagination, the only problem is, the 980Ti really upset the balance of power before the Fury was even launched, as btruner suggested in his excellent editorial opinion article.

The potential to OC the card with such an excellent cooling solution is there, the only thing needed is for tools to be developed to unlock voltage control, then this little card could OC to new heights, dunno if it'll OC as well as Maxwell does, but it will surely go beyond the paltry 50MHz or 75MHz most users and reviewers are getting currently.

That's the untapped potential in Fury, HBM OCing is cool and all that, but as it stands, the card is not starving for memory bandwidth, but that core sure can and will be pushed to the limits by extreme OCers once they get their hands on the right tools, then things will get interesting

Weird. I said all of this (HBM is great tech, Fury - good card, Core needs OC, the HBM is fast enough) but Mirakul thanks your post.... Oh well. Friends grow in mysterious places.

@15th Warlock - the tragedy is I wanted the Fury X to be what it was hyped. I looked forward to buying one (though I'd still need a custom block for it - not wasting 720 worth of rad on one cpu ). I'm still on my 780ti's until blocks come out for the custom cards but if the Fury X get's a bios tweak that allows OC'ing on the core, I'll buy it instead.

 

[Steevo]

This is absolutely not about shitting on AMD. I can see the overclock on the HBM - it's 'unimpressive'. What is holding Fiji back in competitiveness is NOT the very excellent HBM but the very mediocre core that seems so far to be resistant to overclocking. It's amazing that owning a certain brand means you can't comment on another. It's like that statement when parents say "oh, you don't have kids, so you don't know". It's a sham that you can't have s civilised discussion without being labelled troll or fanboy.
I'm a fan of tech but I'm also very aware that as great a new innovation HBM is and FULL kudos to AMD for doing the groundwork to implement it - the card it runs on, no matter how good it is and how well AMD have implemented the cooling system, (see all these fucking positives here?!) - as a techy enthusiast, I want to know I can overclock it to gain MOAR performance. I know a clock for clock Fiji Maxwell fight would probably massively go Fiji's way, especially at 4k.
If people want to argue by saying what I've said that I'm a troll or fanboy then you can go

Spoiler

to a peaceful place, reflect on your life and come to terms with being a better person in such a way as you will find

yourself.

FTR, Mirakul IS a blatant troll. Posts a thread, puts a point across, point is dissected by many - throws hissy fit and calls everyone fanboys or trolls. Steevo - you should know better, unless that was sarcasm.

I visited the peaceful place. It was as good as advised.


Let's have a real speculative conversation on the interposer layer being the reason we can't clock the core very high. Perhaps power delivery is an issue, or the die temp ramps up quickly and the interposer is acting as an insulation layer, or the amount of power being pushed through the traces is close to the maximum, or as suggested the ability to create a larger or more dense interface was the reason it's so lame.

 

[mirakul]

Just as I suspected, the gains aren't anywhere near the big jump in performance alleged to in the previous thread, and more in line with what you would expect from OCing memory on any card.

A bit off topic, but related nontheless, I was just done reading anandtech's excellent review of the Fury, it's an excellent read and I recommend it, after reading it I have a better grasp of how AMD over engineered the Fury, everything from the 500W cooling solution, to the amazing feat of developing the right tools to bring HBM to the masses, seems like AMD threw all of its weight on the Fury, and they can only be commended for doing so.

The Fury is not a bad card by any stretch of the imagination, the only problem is, the 980Ti really upset the balance of power before the Fury was even launched, as btruner suggested in his excellent editorial opinion article.

The potential to OC the card with such an excellent cooling solution is there, the only thing needed is for tools to be developed to unlock voltage control, then this little card could OC to new heights, dunno if it'll OC as well as Maxwell does, but it will surely go beyond the paltry 50MHz or 75MHz most users and reviewers are getting currently.

That's the untapped potential in Fury, HBM OCing is cool and all that, but as it stands, the card is not starving for memory bandwidth, but that core sure can and will be pushed to the limits by extreme OCers once they get their hands on the right tools, then things will get interesting

Yes, the main point is pushing to the limit.
By the way another guy just broke 20k Graphic score with FuryX@630MHz HBM.

Take that with a pinch of salt.

 

[Tatty_Two]

I agree, it was initially said or thought it was locked, it's good that you and others brought that to our attention through these threads, however your initial claims in the original thread were around "large" gains, personally looking at the exact evidence you provided I see very small gains from overclocking the memory with possibly one or two exceptions, now before you get all defensive on me like you have with one or two others, I am not criticising the card, you or anyone, just stating the facts as I see them, which is the very reason why we post stuff like this isn't it? I mean, you saw the evidence before we did, you must have seen that there are gains but they are small, if you didn't want people here to say the gains were small why post?

This is an open forum where we like to say what we think and say what we mean, some don't necessarily always do it in the right manner but most say it as they see it, if you think that's wrong then don't post stuff like this, your contribution in posting this is most welcome but stop being so defensive otherwise it's you that will cause this one to be closed, ohhhh by the way, before you say it....... I'm on the red team! Lets have a healthy discussion without the drama

 

[erocker]

Ok now you are ignoring the game bench, good job on selective reading.

I just wondering, this is Amd box, why ppl without any Amd cards keep coming here to throw flamebait. Have a life, please.

Because this is an open forum where people can post where they please. If you want to post facts on this, overclock your Fury card and post your results. Circumventing a thread closed by a moderator here is not the way to do it.

Oh, I'm on the yellowish-brown team... That's what you get when you mix red and green.

 

[the54thvoid]

ohhhh by the way, before you say it....... I'm on the red team! Lets have a healthy discussion without the drama

Well, nobody's perfect

 

[R-T-B]

I was on the red team... With a R9 290x. Not a bad card, but unlike Nvidia Fiji did not offer a huge change in efficiency to win the day. Maxwell is just better IMO. I think AMDs best competetive move they can make now is price slashing...

Yes, I drank the green koolaid, but it's good stuff, damnit.

 

[Pill Monster]

Yes, the main point is pushing to the limit.
By the way another guy just broke 20k Graphic score with FuryX@630MHz HBM.

Take that with a pinch of salt.

Can you stop trolling man? That's a screenshot of your own system.
Would you like me to post links to all your dupe accounts on OCN, Guru, and Anandtech, along with screenshots? If you think I'm bluffing, I know your old card was a GTX970 (or 980).


This is as bad as the sockpuppets on Guru who compare W7/W8 W10 FPS/benchmark results and claim the increased performance in W10 is because AMD devs have somehow magically optimized their drivers.
In actual fact it's due to the W10 WDM reducing kernel/user mode swapping therefore reducing CPU load, as well as memory management shifting to the application.


Btw my system is AMD, not that it makes a difference.

 

[Mussels]

So this has proven what the non-trolls realised in the other thread: GPU memory bandwidth goes up, but current titles dont need it.


I'd like to see how much of an improvement it makes in crossfire, where you ahve double the GPU power but the same memory bandwidth (since its not additive)

 

[Pill Monster]

So this has proven what the non-trolls realised in the other thread: GPU memory bandwidth goes up, but current titles dont need it.


I'd like to see how much of an improvement it makes in crossfire, where you ahve double the GPU power but the same memory bandwidth (since its not additive)

Now that would be a topic worth discussing, I'd like to know myself tbh.
Since W10 GPU memory is meant to be virtualized multiple GPU's could share the available VRAM....(kinda like having 2 GPU's on 1 PCB). In that scenario PCI-e will eventually become the bottleneck (say with Quadfire for example)....

 

[namae_nanka]

Gains seems bigger at 1080p(and perhaps win7) and of course depend on the application used. Kombustor showed substantial increase with HBM while only 1fps with core increase.

http://www.overclock.net/t/1547314/...nano-x-x2-fiji-owners-club/1720#post_24106947

The highest graphics scores in firestrike without tesselation modification are around 17.6k.

 

[EarthDog]

Just as I suspected, the gains aren't anywhere near the big jump in performance alleged to in the previous thread, and more in line with what you would expect from OCing memory on any card.

took two threads to confirm the point we all made in the original or wanted to get down to in the original... nice.

In the end, there is no magic above and beyond typical memory scaling. No magic sweetspot like was claimed earlier....

But as it stands the performance gains demonstrated by your example's DO NOT MATCH THE HYPE

Fitting though, isn't it? AMD is a marketing machine, but in the end, a lot of users, including from team red, are left disappointed because of it. I wish they would slow down a but on the hype and speed up a bit on substance (r9 295x2 owner for the record).

 

[Aquinus]

I'd like to see how much of an improvement it makes in crossfire, where you ahve double the GPU power but the same memory bandwidth (since its not additive)

Don't confuse bandwidth with capacity. Currently GPUs might have to duplicate video memory but keep in mind that both GPUs are probably doing the same task, just sharing the load. If you don't get any gains in memory bandwidth with a single GPU, I seriously doubt you will with two because it's just a second GPU doing the same calculations (assuming AFR is being used.) The only exception to that rule is if you have more GPU compute than VRAM that it can utilize which turns into the streaming textures problem.

In all candor, I think it's pretty clear that HBM is not the bottleneck and that overclocking it definitely shows diminishing returns from the get go whereas overclocking the core shows very tangible results, just that you can't push it very far.

The OP seems to avoid all of the facts that are contrary to his view, which seems ass backwards and straight up biased. Maybe I should sit back, make some popcorn, and watch how many people tell off the OP.

 

[EarthDog]

Don't confuse bandwidth with capacity. Currently GPUs might have to duplicate video memory but keep in mind that both GPUs are probably doing the same task, just sharing the load. If you don't get any gains in memory bandwidth with a single GPU, I seriously doubt you will with two because it's just a second GPU doing the same calculations (assuming AFR is being used.) The only exception to that rule is if you have more GPU compute than VRAM that it can utilize which turns into the streaming textures problem.

In all candor, I think it's pretty clear that HBM is not the bottleneck and that overclocking it definitely shows diminishing returns from the get go whereas overclocking the core shows very tangible results, just that you can't push it very far.

The OP seems to avoid all of the facts that are contrary to his view, which seems ass backwards and straight up biased. Maybe I should sit back, make some popcorn, and watch how many people tell off the OP.

quoting for truth/permanence.

 

[dorsetknob]

The OP seems to avoid all of the facts that are contrary to his view, which seems ass backwards and straight up biased. Maybe I should sit back, make some popcorn, and watch how many people tell off the OP.

as in threads previous started as well as this one also Sad to say probably in threads yet to be started because this one is not going the way the OP wants it to go

 

[uuuaaaaaa]

We need to find someone that is kind enough to do some extensive testing with a Fury X that does 625 MHz on memory.

 

[mirakul]

Waiting for FuryX owners though. It is the purpose of this thread by the way.

 

[MxPhenom 216]

Waiting for FuryX owners though. It is the purpose of this thread by the way.

I dont think there is a single Fury X owner on TPU.

 

[mirakul]

There are more than one @v12dock
http://www.techpowerup.com/forums/threads/overclocked-hbm-its-true-and-its-fast.213875/#post-3306403

 

[Assimilator]

AMD did NOT go with HBM because Fiji is bandwidth-starved. They went for HBM because its controllers use a lot less power than GDDR5, and then put that saved power back into the core. Since Fiji couldn't add any more ROPs than Hawaii, the only way they could use the extra juice was by pushing the core up to high frequencies. A high-clocked core would put out a lot of heat, hence the over-engineered cooling solution.

But for whatever reason - perhaps the newness of HBM/interposer - AMD weren't able to hit the core speed they were aiming for. If, as I suspect, it's a fundamental issue with the silicon, then no BIOS or driver updates or game optimisations are going to make a difference.

We need to find someone that is kind enough to do some extensive testing with a Fury X that does 625 MHz on memory.

Why? So we can see the benchmark scores go up by another insignificant amount?

 

[uuuaaaaaa]

AMD did NOT go with HBM because Fiji is bandwidth-starved. They went for HBM because its controllers use a lot less power than GDDR5, and then put that saved power back into the core. Since Fiji couldn't add any more ROPs than Hawaii, the only way they could use the extra juice was by pushing the core up to high frequencies. A high-clocked core would put out a lot of heat, hence the over-engineered cooling solution.

But for whatever reason - perhaps the newness of HBM/interposer - AMD weren't able to hit the core speed they were aiming for. If, as I suspect, it's a fundamental issue with the silicon, then no BIOS or driver updates or game optimisations are going to make a difference.



Why? So we can see the benchmark scores go up by another insignificant amount?

To confirm or debunk OP's claims doing the experiment, that's it.

 

[HTC]

To confirm or debunk OP's claims doing the experiment, that's it.

Why not underclocking instead of overclocking? All it's needed is the % difference VS stock and it can be done by going over or going under, unless ofc such low clocks aren't permitted by the card (doesn't function properly).

 

[uuuaaaaaa]

Why not underclocking instead of overclocking? All it's needed is the % difference VS stock and it can be done by going over or going under, unless ofc such low clocks aren't permitted by the card (doesn't function properly).

OP claimed that the card would have a sudden non-linear performance increase around the 620MHz range.